Sometimes Manual Single Node Transaction Recovery is Required with HornetQ

When killing/restarting a single HornetQ node during message consumption, messages go missing. They are pending however (DeliveringCount=# of Missing Messages). These messages are also visible in the PreparedTransactions table. Waiting a long time does not affect the pending status, and neither does restarting the server. They are never recovered. Manually rollingback/committing the transaction does seem to resolve this situation. Why should that be necessary?